作者：向瑾，付广，张静，张晓宁，张劭

单位：南华大学附属第一医院胃肠外科，湖南 衡阳 421001

Authors: Xiang Jin, Fu Guang, Zhang Jing, Zhang Xiaoning, Zhang Shao

Unit: Department of Gastrointestinal Surgery, the First Affiliated Hospital of South China University, Hengyang 421001, Hunan, China

摘要：

目的 利用增强螺旋计算机断层扫描（computed tomography, CT）影像数据，探讨卷积神经网络深度学习模型在胃左动脉变异检测方面的可行性。方法 回顾性选取2019年1月至12月在南华大学附属第一医院行增强螺旋CT的305例门诊患者的影像学资料。通过医生阅读患者的增强CT血管图像，对变异胃周动脉进行分类并标注。将所有数据随机分为五组，4个训练组，1个测试组。构建分类-检测级联框架模型对数据进行深度学习，计算平均曲线下面积（area under the curve, AUC）、查全率、查准率和准确率评估该模型的性能。结果 共39例患者存在胃左动脉变异血管，胃左动脉变异发生率约12.8%。最常见的2种变异类型是胃左动脉发出替代肝左动脉（12/305，3.9%）和副肝左动脉（13/305，4.3%），而胃左动脉缺如的现象比较罕见（2/305，0.7%）。分类网络深度学习模型五组的平均AUC、查全率、查准率、准确率分别为0.82、73.3%、78.2%、79.0%，检测网络深度学习模型五组的平均AUC、查全率、查准率、准确率分别为0.87、65.6%、87.7%、77.8%。结论 与胃左动脉相关的变异血管中，替代/副肝左动脉最为常见。构建的卷积神经网络深度学习模型具有较好的胃左动脉变异检测效能。

关键词： 深度学习；血管变异；胃左动脉；肝总动脉；替代/副肝左动脉

Abstract：

Objective  To explore the feasibility of using a convolutional neural network deep learning model to detect variations in the left gastric artery by utilizing enhanced spiral computed tomography (CT) imaging data. Method  305 outpatient cases who underwent enhanced spiral CT scanning from January to December 2019 at the First Affiliated Hospital of South China University were retrospectively selected. The doctors read the enhanced CT angiogram images of the patients to classify and label variants of the left gastric artery. All data were randomly divided into five groups: four training groups and one testing group. A classification-detection cascaded framework model was constructed to perform deep learning on the data and obtain the average area under the curve (AUC), recall, precision and accuracy to evaluate the performance of the model. Result  There were 39 cases of variation in blood vessels related to the left gastric artery, with an incidence rate of approximately 12.8%. The two most common types of variation were the left gastric artery branching off to the replaced left hepatic artery (12/305, 3.9%) and the accessory left hepatic artery (13/305, 4.3%). Absence of the left gastric artery was relatively rare (2/305, 0.7%). The average AUC, recall, precision and accuracy of the classification network deep learning model for the five groups were 0.82, 73.3%, 78.2% and 79.0%, respectively. The average AUC, recall, precision and accuracy of the detection network deep learning model for the five groups were 0.87, 65.6%, 87.7%and 77.8%, respectively. Conclusion  The replaced/accessory left hepatic artery is the most common variation related to the left gastric artery. The constructed convolutional neural network deep learning model has good performance in detecting variations in the left gastric artery.

Key Words: Deep learning; Vascular variation; Left gastric artery; Common hepatic artery; Replaced/accessory left hepatic artery